Lendek 2010

Fuzzy Sets and Systems 161 (2010) 2043 – 2065
www.elsevier.com/locate/fss
Adaptive observers for TS fuzzy systems with

unknown polynomial inputs
Zs. Lendeka,∗ , J. Lauberb , T.M. Guerrab , R. Babuškaa , B. De Schuttera, c
a Delft Center for Systems and Control, Delft University of Technology, Mekelweg 2, 2628 CD Delft, The Netherlands
b Université de Valenciennes, BP 311, Le Mont Houy, 59313 Valenciennes Cedex 9, France
c Marine and Transport Technology Department, Delft University of Technology, The Netherlands
Received 23 September 2009; received in revised form 9 March 2010; accepted 16 March 2010
Available online 20 March 2010
Abstract
A large class of nonlinear systems can be well approximated by Takagi–Sugeno (TS) fuzzy models, with linear or affine consequents.
However, in practical applications, the process under consideration may be affected by unknown inputs, such as disturbances, faults
or unmodeled dynamics. In this paper, we consider the problem of simultaneously estimating the state and unknown inputs in TS
systems. The inputs considered in this paper are (1) polynomials in time (such as a bias in the model or an unknown ramp input
acting on the model) and (2) unmodeled dynamics. The proposed observer is designed based on the known part of the fuzzy model.
Conditions on the asymptotic convergence of the observer are presented and the design guarantees an ultimate bound on the error
signal. The results are illustrated on a simulation example.
© 2010 Elsevier B.V. All rights reserved.
Keywords: TS fuzzy systems; Fuzzy observers; Lyapunov stability; Adaptive observers; Unknown input observers
1. Introduction
Many problems in decision making, control, and monitoring require the estimation of states and possibly uncertain
parameters, based on a dynamic system model and a sequence of noisy measurements. For such a purpose, dynamic
systems are often modeled in the state-space framework, using the state-transition model, which describes the evolution
of states over time, and the sensor model, which relates the measurements to the states.
A generic method for the design of an observer valid for all types of nonlinear systems has not yet been developed.
Moreover, a large class of nonlinear systems can be well approximated by TS fuzzy models [1], which in theory can
approximate a general nonlinear system to an arbitrary degree of accuracy [2]. Stability conditions have been derived
for TS fuzzy systems, most of them relying on the feasibility of an associated system of linear matrix inequalities
(LMIs) [3–5]. A comprehensive survey on the analysis of fuzzy systems can be found in [6].
For a TS fuzzy model, well-established methods and algorithms can be used to design observers that estimate
unmeasurable states. Several types of such observers have been developed for TS fuzzy systems, among which: fuzzy
∗ Corresponding author. Tel.: +31 15 27 88573; fax: +31 15 27 86679.

E-mail addresses: z.lendek@tudelft.nl (Zs. Lendek), jimmy.lauber@univ-valenciennes.fr (J. Lauber), thierry.guerra@univ-valenciennes.fr
(T.M. Guerra), r.babuska@tudelft.nl (R. Babuška), b.deschutter@tudelft.nl (B. De Schutter).
0165-0114/$ - see front matter © 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.fss.2010.03.010
2044 Zs. Lendek et al. / Fuzzy Sets and Systems 161 (2010) 2043 – 2065
Thau–Luenberger observers [3,7], reduced-order observers [5,8], and sliding-mode observers [9]. In general, the design
methods for observers also lead to an LMI feasibility problem.
Adaptive observers are observers that simultaneously estimate the states and unknown parameters, by processing
the measurements online. For SISO LTI systems, the adaptive observer design has been largely investigated (see [10]
and the references therein).
A general tendency in case of nonlinear systems, both in adaptive controller and observer design [11–15], is to
assume that the system is SISO and in observer canonical form. By using a quadratic Lyapunov function, ensuring
strictly positive real conditions, the Kalman–Yakubovic–Popov lemma is applied and the adaptive laws are deduced
from the Lyapunov synthesis. A shortcoming of these observers is that they do not incorporate prior information (such
as an approximate model) and that they cannot be used when physical states have to be estimated, or when a model
is not in a canonical form. Robust versions of these adaptive observers have also been derived for systems affected
by a bounded disturbance, by adding a robustness term [11,13,15–17]. In many cases, when using both an observer
and a controller, the robustness term is incorporated in the controller instead, to deal with the approximation error
and disturbances [12,14,18,19], amended in [20]. Results for MIMO systems include high gain observers [21], special
observer canonical forms of the system [22,23], linearly parameterized neural networks [24,25] and observers based
on a known linear part of the model [26]. A linear observer applied for a nonlinear system in general can be used only
in a small neighborhood of the linearization point. Moreover, such an observer can only deal with constant or slowly
varying inputs [27].
In this paper, we consider polynomial unknown inputs (such as a bias in the model or an unknown ramp input acting on
the model) and uncertainty in the model dynamics. Model-based observer design in the presence of bounded unknown
inputs has been investigated recently [28–32]. However, these approaches estimate only the states, and attenuate the
effect of the disturbances, but do not estimate the unknown inputs. Our objective is to estimate both the state and the
unknown input from the available input and output information. The design of observers to estimate unknown inputs
is an important problem, since in many cases not all the inputs are known [27,31,32], but is needed for proper planning
and control. For instance, in machine tool and manipulator applications, the cutting force exerted by the tool or the
exerting force/torque of the robot is of interest, but it is very difficult or expensive to measure [26,33]. Load estimation
in e.g., electricity distribution networks [34], or wind turbines [35] is necessary for proper planning and operation. In
biomechanics, the myoskeletal system can be regarded as a dynamic system, where segment positions and trajectories
are the system outputs and joint torques are the nonmeasurable inputs [36]. In traffic control, time-varying parameters
have to be tracked, which can be regarded as unknown inputs [37]. In chaotic systems, for chaos synchronization and
secure communication, one has to estimate not only the state, but also the input signal [38]. The class of adaptive
observers has received considerable interest in fault detection and identification, where the unknown inputs represent
the effect of actuator failures or plant components and its presence has to be detected as soon as possible. However,
many of these methods only detect the fault or attenuate its effect on the states [39], and rarely reconstruct it [40].
We develop a method to design observers for TS fuzzy systems with polynomial unknown inputs or unmodeled
dynamics. The idea behind this type of design is that, in practice, a TS representation of a nonlinear system may
be obtained by identifying LTI models in different operation points. However, such models are often inaccurate. By
estimating the unknown inputs or unmodeled dynamics, the accuracy of the identified models can be improved. The
observers are designed based on the already identified model, such that, together with the appropriate update law
estimating the unknown inputs, they ensure the convergence of the estimation error.
The structure of the paper is as follows. Section 2 reviews stability conditions for TS fuzzy systems and observers.
Section 3 presents the model and the observer considered. The proposed design for observers in the case of polynomial
unknown inputs is described in Section 4. Section 5 presents the observer for estimating unmodeled dynamics. Related
work is presented in Section 6. An example is given in Section 7. Finally, Section 8 concludes the paper.
2. Stability of fuzzy systems and observers
Though the consequents of dynamic TS fuzzy systems are usually chosen linear or affine, it is well-known that the
stability of these local models does not imply the stability of the overall fuzzy system. In the full fuzzy model, the local
models are blended, and the particular blending of several local models may be strongly nonlinear, which influences
the stability. Several stability conditions have been derived to ensure the stability of the system for any normalized
membership function, independent of the operators used in the antecedent rules [3]. Most of these conditions depend
Zs. Lendek et al. / Fuzzy Sets and Systems 161 (2010) 2043 – 2065 2045
on the feasibility of an associated LMI problem. In this section, we introduce the notations used throughout the paper
and review some conditions for the stability of fuzzy systems and observer design. Throughout the paper it is assumed
that the membership functions of the TS system are normalized.
2.1. Preliminaries
Throughout the paper I denotes the identity matrix of the appropriate dimension, H( A) denotes the Hermitian of the
matrix A, i.e., H(A) = A + A T .
Definition 1 (Uniform ultimate boundedness, Khalil [41]). The solutions of ẋ = f(t, x) are uniformly ultimately
bounded with ultimate bound b if there exist b, c > 0 and for every 0 < a < c, there exists T = T (a, b) ≥ 0
such that
x(t0 ) ≤ a ⇒ x(t) ≤ b ∀t ≥ t0 + T
2.2. Autonomous fuzzy systems
Consider the autonomous fuzzy system:

m
ẋ = wi (z)Ai x (1)
i=1
where Ai , i = 1, 2, . . . , m represents the ith local linear model, wi is the corresponding normalized membership
function, and z the vector of the scheduling variables.
For system (1), several stability conditions have been derived. Among them, a well-known and frequently used
condition is formulated below [3].
Theorem 1. System (1) is exponentially stable if there exists P = P T > 0 so that H(P Ai ) < 0, for i = 1, 2, . . . , m.
Remark. Less conservative, but more complex conditions can be obtained by using relaxations for LMIs [3,6,42,43].
A condition on the convergence rate of system (1) was also derived from Theorem 1 [3].
Theorem 2. The decay rate of system (1) is at least , if there exists P = P T > 0, so that
H(P Ai ) + 2P < 0, i = 1, 2, . . . , m
Stability conditions similar to those of Theorem 1 can be used if the system considered is subjected to disturbances.
Consider the following perturbed fuzzy system:

m
ẋ = wi (z)Ai x + D f (t, x) (2)
i=1
where D is a perturbation distribution matrix and f is Lipschitz in x, i.e., there exists > 0 so that f(t, x) ≤ x, for
all t and x. With these assumptions, a sufficient stability condition can be formalized by the following theorem [42].
Theorem 3. System (2) is exponentially stable if there exist matrices P = P T , Q = Q T , so that

P>0 Q>0
min (Q)
≤
P D
H(P Ai ) < −2Q, i = 1, 2, . . . , m (3)
where min is the smallest eigenvalue.
Several variants of the above theorem exist, together with algorithms to compute robustness measures [42].
Other approaches, based on partitioning the state-space into regions determined by simultaneously active membership
functions can be found in [4,44,45]. Similar conditions for the discrete-time case are described in [46].
2.3. Fuzzy observers
Consider now the affine fuzzy system

m
ẋ = wi (z)(Ai x + Bi u + ai )
i=1

m
y= wi (z)Ci x (4)
i=1
and an observer of the form

m
x̂˙ = wi (ẑ)(Ai x̂ + Bi u + ai + L i (y − ŷ))
i=1

m
ŷ = wi (ẑ)Ci x̂ (5)
i=1
As before, it is assumed that the membership functions are normalized. Depending on the explicit form of the error
˙ the theorems presented in Section 2.2 can be directly applied, or similar conditions may
dynamics given by ė = ẋ − x̂,
be derived to ensure the convergence to zero of the estimation error. For the design, two cases are distinguished: (1)
the scheduling vector z does not depend on the estimated states, i.e., ẑ = z and (2) ẑ depends on (some of) the states to
be estimated.
2.3.1. The scheduling vector does not depend on the estimated states
In this case, the error system can be written as

m
m
ė = wi (z)w j (z)( Ai − L i C j )e
i=1 j=1

m
m
m
= wi (z)wi (z)(Ai − L i Ci )e + wi (z)w j (z)( Ai − L i C j + A j − L j Ci )e (6)
i=1 i=1 j=1
j>i
Using a Lyapunov function of the form V (t) = eT Pe, with P = P T > 0, basic sufficient stability conditions for
this system were derived in [3]:
Theorem 4. The system (6) is GAS, if there exists P = P T > 0 so that for i = 1, 2, . . . , m and j = 1, 2, . . . , m:
H(P( Ai − L i C j + A j − L j Ci )) < 0 ∀ j ≥ i : ∃ z : wi (z)w j (z) 0 (7)
A well-known condition on the design of the observer for the system (1), so that a desired convergence rate is
guaranteed, has also been obtained in [3] by appropriately modifying Theorem 2.
2.3.2. The scheduling vector depends on the estimated states

The second case is when the scheduling vector depends on the states to be estimated. For the simplicity of the
notation, only the case with common measurement matrices (Ci = C, i = 1, 2, . . . , m) is presented. Then, the
observer (5) becomes

m
x̂˙ = wi (ẑ)(Ai x̂ + Bi u + ai + L i (y − ŷ))
i=1
ŷ = C x̂ (8)
and the error dynamics can be expressed as

m
m
ė = wi (ẑ)(Ai − L i C)e + (wi (z) − wi (ẑ))(Ai x + Bi u + ai ) (9)
i=1 i=1
For such a system, sufficient stability conditions are given by the following theorem [42].
Theorem 5. The error dynamics (9) are exponentially stable, if there exist > 0, P = P T > 0, Q = Q T > 0 so that
for all i = 1, . . . , m
H(P(Ai − L i C)) ≤ −Q

Q − 2 I P
>0
P I
m

(wi (z) − wi (ẑ))(Ai x + Bi u + ai ) ≤ e (10)

i=1
m
i.e., i=1 (wi (z) − wi (ẑ))(Ai x + Bi u + ai ) is Lipschitz in e.
Note that since fuzzy systems are generally defined on a compact set, the condition that “there exists > 0 so that
m
i=1 (wi (z) − wi (ẑ))( Ai x + Bi u + ai ) ≤ e” is satisfied if wi (z), i = 1, 2, . . . , m are differentiable w.r.t. x almost
everywhere and have a bounded first derivative for almost all x. In practice, most membership functions used satisfy
this condition, which is therefore not restrictive.
3. Problem statement
Consider the following fuzzy system:

m
ẋ = wi (z)(Ai x + Bi u + ai + Mi d)
i=1

m
y= wi (z)Ci x (11)
i=1
where Ai , Bi , Mi , ai , Ci , i = 1, 2, . . . , m are known and the vector d is an unknown input. Design of unknown input
observers is very important in practice, since, in many cases, not all the inputs are known. Moreover, these inputs
can represent disturbances acting on the process, effects of uncertain dynamics or actuator failures. In this paper we
consider two types of unknown inputs: (1) polynomials in time and (2) unmodeled dynamics of some of the states. Our
goal is to design a stable observer to estimate simultaneously both the state vector x and the unknown inputs d.
The observer considered is of the form

m
x̂˙ = wi (ẑ)(Ai x̂ + Bi u + ai + Mi dˆ + L i (y − ŷ))
i=1

m
ŷ = wi (ẑ)Ci x̂
i=1
˙ ˆ w(ẑ), x̂, y)
d̂ = f (d, (12)
where L i , i = 1, 2, . . . , m, are the gain matrices to be designed for each rule, and f , the update law for d, should be
determined so that the estimation errors x − x̂ and d − dˆ converge asymptotically to zero.
Two main cases can be distinguished, depending on whether or not the scheduling vector depends on the states to
be estimated. The observer design is considered in both cases.
4. Polynomial unknown inputs
In this section, we consider the case when the unknown input is or can be approximated by a polynomial function in
time. Such inputs may represent biases in the model, time-varying disturbances acting on the process, or the degradation
in time or even failure of actuators. We give conditions to design a fuzzy observer and we also give a bound on the
estimation error.
4.1. Observer design
Consider the TS fuzzy system of the form

m
ẋ = wi (z)(Ai x + Bi u + Mi d + ai )
i=1

m
y= wi (z)Ci x (13)
i=1
where there exists p ∈ N so that d( p) = 0, i.e., the unknown input is a p−1-th order polynomial in time. It is assumed
that the states, the unknown input d, and the derivatives of d are observable from y.
4.1.1. Case 1
If z does not depend on x̂, then
Theorem 6. The estimation error using the observer

m
x̂˙ = wi (z)[Ai x̂ + Bi u + L i (y − ŷ) + Mi dˆ + ai ]
i=1

m
ŷ = wi (z)Ci x̂
i=1
( p)
m
dˆ =
p
wi (z)i (y − ŷ)
i=1
(k)
m
(k+1)
dˆ = wi (z)(ik (y − ŷ) + dˆ )
i=1
for k = 1, . . . , p − 1 (14)
is exponentially stable if there exist P = > 0, PT L i , ik ,
i = 1, 2, . . . , m, j = 1, 2, . . . , m, k = 1, 2, . . . , p so that
⎛ ⎛ ⎞⎞
A i − L i C j + A j − L j C i Mi + M j 0 · · · 0
⎜ ⎜ ⎟⎟
⎜ ⎜ −i1 C j − 1j Ci 0 2I · · · 0 ⎟⎟
⎜ ⎜ ⎟⎟
⎜ ⎜ .. .. .. . . .. ⎟ ⎟
⎜
H ⎜P ⎜ ⎜ . . ⎟ ⎟
. . . ⎟⎟ < 0
⎜ ⎜ ⎟ ⎟
⎜ ⎜ − p−1 C j − p−1 Ci 0 0 · · · 2I ⎟ ⎟
⎝ ⎝ i j ⎠⎠
p p
−i C j − j Ci 0 0 ··· 0
∀ j ≥ i : ∃ z : wi (z)w j (z) 0 (15)
Proof. An extended error system, containing both the state error and the derivatives of the input error d̄ = d − d,ˆ can
be expressed as
⎛ ⎞ ⎛ ⎞
ė A i − L i C i Mi 0 · · · 0 ⎛ ⎞
⎜ ˙ ⎟ ⎜ −1 C e
⎜ d̄ ⎟
⎜ ⎟ ⎜ i i 0 I ··· 0 ⎟ ⎟⎜ ⎟
⎜ ¨ ⎟ m ⎜ ⎟⎜ d̄ ⎟
⎜ . . . .
.. .. . . . .. ⎟ ⎜ ⎟
ėa = ⎜
⎜
d̄ ⎟ =
⎟ wi (z)wi (z) ⎜ .. ⎟⎜ .. ⎟
⎜ . ⎟ i=1 ⎜ ⎟ ⎜ . ⎟
⎜ .. ⎟ ⎜ p−1 ⎟⎝ ⎠
⎝ ⎠ ⎝ −i Ci 0 0 · · · I ⎠ ( p−1)
p d̄
( p)
d̄ −i Ci 0 0 ··· 0
⎛ ⎞
A i − L i C j + A j − L j C i Mi + M j 0 ··· 0 ⎛ ⎞
⎜ ⎟ e
⎜ −i1 C j − 1j Ci 0 2I · · · 0 ⎟⎜ ⎟
⎜ ⎟⎜ d̄ ⎟
m m
⎜ .. .. .. . . .. ⎟⎜ ⎟
+ wi (z) w j (z) ⎜
⎜ . . . . . ⎟⎜
⎟⎜ .. ⎟ (16)
⎜ ⎟⎝ . ⎟
i=1 j=1 ⎜ p−1 p−1 ⎟ ⎠
j>i ⎝ −i C j − j Ci 0 0 · · · 2I ⎠ ( p−1)
p p d̄
−i C j − j Ci 0 0 ··· 0
Using a quadratic Lyapunov function for the extended error vector V = eaT Pea , the derivative is expressed as
⎛ ⎛ ⎞⎞
A i − L i C i Mi 0 · · · 0
⎜ ⎜ −1 C I ··· 0⎟ ⎟
⎜ ⎜ i i 0 ⎟⎟

m ⎜ ⎜ ⎟⎟
⎜ ⎜ .. .. .. . . .. ⎟⎟
V̇ = wi (z)wi (z)ea H ⎜ P ⎜
T
. . . ..⎟ ⎟ ea
⎜ ⎜ ⎟⎟
i=1 ⎜ ⎜ p−1 ⎟⎟
⎝ ⎝ −i Ci 0 0 · · · I ⎠⎠
p
−i Ci 0 0 ··· 0
⎛ ⎛ ⎞⎞
A i − L i C j + A j − L j C i Mi + M j 0 ··· 0
⎜ ⎜ ⎟⎟
⎜ ⎜ −i1 C j − 1j Ci 0 2I · · · 0 ⎟⎟
⎜ ⎜ ⎟⎟
m m
⎜ ⎜ .. .. .. . . ..⎟⎟
+ wi (z) w j (z)eaT H ⎜ ⎜
⎜P ⎜ . . . . .⎟⎟ ea
⎟⎟
⎜ ⎜ ⎟⎟
i=1 j=1
⎜ ⎜ −i
p−1
Cj − j
p−1
Ci 0 0 · · · 2I ⎟ ⎟
j>i
⎝ ⎝ ⎠⎠
p p
−i C j − j Ci 0 0 ··· 0
which is negative definite if the condition (15) is satisfied.
Remark. In order to design observers with a desired convergence rate , Theorem 2 can be applied. The estimation
error of the observer (14) converges with a rate at least if there exists P = P T > 0, L i , ik , i = 1, 2, . . . , m,
j = 1, 2, . . . , m, k = 1, 2, . . . , p so that
⎛ ⎛ ⎞⎞
⎜ ⎜ ⎟⎟
⎜ ⎜ −i1 C j − 1j Ci 0 2I · · · 0 ⎟⎟
⎜ ⎜ ⎟⎟
⎜ ⎜ .. .. .. . . .. ⎟ ⎟
⎜
H ⎜P ⎜ ⎜ . . ⎟ ⎟
. . . ⎟⎟ + 4P < 0
⎜ ⎜ ⎟ ⎟
⎜ ⎜ − p−1 C j − p−1 Ci 0 0 · · · 2I ⎟ ⎟
⎝ ⎝ i j ⎠⎠
p p
−i C j − j Ci 0 0 ··· 0
∀ j ≥ i : ∃ z : wi (z)w j (z) 0 (17)
The proof follows directly.

4.1.2. Case 2
A similar observer can also be designed if z depends on x̂. For the simplicity of the notation, only the case when the
measurement matrix is common for all rules is presented. Note, however, that if the measurement matrices are different
for the rules, the observer can be designed analogously.
The observer considered is of the form

m
x̂˙ = wi (ẑ)[Ai x̂ + Bi u + L i (y − ŷ) + Mi dˆ + ai ]
i=1
ŷ = C x̂
( p)
m
dˆ =
p
wi (ẑ)i (y − ŷ)
i=1
( j)
m
( j+1)
dˆ = wi (ẑ)(i (y − ŷ) + dˆ
j
)
i=1
for j = 1, . . . , p − 1 (18)
Then the extended error system becomes

⎛ ⎞
A i − L i C Mi 0 · · · 0 ⎛ ⎞
⎜ −1 C ⎟ e
⎜ i 0 I ··· 0 ⎟⎜ ⎟
m ⎜ ⎟⎜ d̄ ⎟
⎜ .. .. .. . . .. ⎟ ⎜ ⎟
ėa = wi (ẑ) ⎜ . . . . . ⎟⎜ .. ⎟
⎜ ⎟⎜ . ⎟
i=1 ⎜ p−1 ⎟⎝ ⎠
⎝ −i C 0 0 ··· I ⎠ ( p−1)
p d̄
−i C 0 0 ··· 0

m
+ (wi (z) − wi (ẑ))(I 0 . . . 0)T · (Ai x + Bi u + Mi d + ai ) (19)
i=1
m
Assuming that i=1 (wi (z) − wi (ẑ))(Ai x + Bi u + Mi d + ai ) ≤ e, and therefore it is also Lipschitz in ea , with
the same Lipschitz constant, the stability conditions become:
j
Corollary 1. The error system (19) is asymptotically stable, if there exist P = P T > 0, Q = Q T > 0, L i , i ,
i = 1, 2, . . . , m, j = 1, 2, . . . , p so that
⎛ ⎛ ⎞⎞
A i − L i C Mi 0 · · · 0
⎜ ⎜ −1 C I ··· 0⎟ ⎟
⎜ ⎜ i 0 ⎟⎟
⎜ ⎜ ⎟⎟
⎜ ⎜ .. .. .. . . .. ⎟⎟
H ⎜P ⎜ . . . . .⎟ ⎟ < −Q
⎜ ⎜ ⎟⎟
⎜ ⎜ p−1 ⎟⎟
⎝ ⎝ −i C 0 0 · · · I ⎠⎠
p
−i C 0 0 ··· 0
m

(wi (z) − wi (ẑ))( Ai x + Bi u + Mi d + ai ) ≤ e

i=1

Q − 2 I P
>0 (20)
P I

4.2. Bound on errors
In most cases, the unknown input is not polynomial, but it is possible to determine a bound on some derivative of it.
Therefore, consider the case when there exists p ∈ N so that d( p) is bounded by a known constant, i.e., d( p) < d ,
and d( j) , j = 1, 2, . . . , p are observable from y. In this case, although the estimation error does not converge to zero,
it is bounded, and an upper bound on it can be computed as follows.
4.2.1. Case 1
If the scheduling vector does not depend on x̂, the error system can be written as
⎛ ⎞ ⎛ ⎞
A i − L i C i Mi 0 · · · 0 0
⎜ −1 C 0 I ··· 0 ⎟ ⎜ 0 ⎟
⎜ i i ⎟ ⎜ ⎟
m ⎜ ⎟ ⎜ ⎟
⎜ .. .. .. . . .. ⎟ ⎜ 0 ⎟
ėa = wi (z)wi (z) ⎜ . . . . ⎟
.⎟ ea + ⎜ ⎟
⎜ ⎜ . ⎟
i=1 ⎜ ⎟ ⎜ . ⎟
p−1
⎝ −i Ci 0 0 · · · I ⎠ ⎝ . ⎠
p
−i Ci 0 0 ··· 0 d( p)
⎛ ⎞
⎜ ⎟
⎜ −i1 C j − 1j Ci 0 2I · · · 0 ⎟
⎜ ⎟
m m
⎜ .. .. .. . . ..⎟
+ wi (z) w j (z) ⎜⎜ . . . . .⎟ ea
⎟ (21)
⎜ ⎟
i=1 j=1 ⎜ − p−1 p−1
C j − j Ci 0 0 ··· 2I ⎟
j>i ⎝ i ⎠
p p
−i C j − j Ci 0 0 ··· 0
Theorem 7. The error described by (21) is ultimately bounded by a ball with radius

max (P) max (P)d
= (22)
min (P) min (Q)
if there exist P = P T > 0, Q = Q T > 0, L i , ik , i = 1, 2, . . . , m, k = 1, 2, . . . , p, j = 1, 2, . . . , m so that

d( p) < d
⎛ ⎛ ⎞⎞
⎜ ⎜ ⎟⎟
⎜ ⎜ −i1 C j − 1j Ci 0 2I · · · 0 ⎟⎟
⎜ ⎜ ⎟⎟
⎜ ⎜ .. .. .. . . .. ⎟ ⎟
⎜
H ⎜P ⎜ ⎜ . . ⎟ ⎟
. . . ⎟⎟ < −4Q
⎜ ⎜ ⎟ ⎟
⎜ ⎜ − p−1 C j − p−1 Ci 0 0 · · · 2I ⎟⎟
⎝ ⎝ i j ⎠⎠
p p
−i C j − j Ci 0 0 ··· 0
∀ j ≥ i : ∃ z : wi (z)w j (z) 0 (23)
for i = 1, 2, . . . , m, where ∈ (0, 1) and min and max denote the eigenvalues with the smallest and largest absolute
magnitude, respectively.
Proof. Consider a quadratic Lyapunov function for the extended error vector, V = eaT Pea . Then,
⎛ ⎛ ⎞⎞
A i − L i C i Mi 0 · · · 0
⎜ ⎜ −1 C 0 I ··· 0 ⎟ ⎟
⎜ ⎜ i i ⎟⎟
m ⎜ ⎜ ⎟⎟
⎜ ⎜ .. .. .. . . .. ⎟⎟
V̇ = wi (z)wi (z)eaT H ⎜ P ⎜ . . . . . ⎟⎟ ea
⎜ ⎜ ⎟⎟
i=1 ⎜ ⎜ p−1 ⎟⎟
⎝ ⎝ −i Ci 0 0 · · · I ⎠⎠
p
−i Ci 0 0 ··· 0
⎛ ⎛ ⎞⎞
A i − L i C j + A j − L j C i Mi + M j 0 ··· 0
⎜ ⎜ ⎟⎟
⎜ ⎜ −i1 C j − 1j Ci 0 2I · · · 0 ⎟⎟
⎜ ⎜ ⎟⎟
m m
⎜ ⎜ .. .. .. . . .. ⎟⎟
+ wi (z) w j (z)ea H ⎜
T ⎜
⎜P ⎜ . . . . . ⎟⎟ ea
⎟⎟
⎜ ⎜ ⎟⎟
i=1 j=1
⎜ ⎜ −i
p−1
Cj − j
p−1
Ci 0 0 · · · 2I ⎟ ⎟
j>i
⎝ ⎝ ⎠⎠
p p
−i C j − j Ci 0 0 ··· 0

m
T
+ wi (z)2eaT P(0 0 0 · · · d( p) )T
i=1
≤ −2min (Q)ea 2 + 2max (P)ea d

≤ −2(1 − )min (Q)ea 2 − 2(min (Q)ea 2 − max (P)ea d )
where ∈ (0, 1) is arbitrarily chosen and Q = Q T is a positive definite matrix such that (23) is satisfied. Then, V̇ is
negative definite if
min (Q)ea 2 − max (P)ea d > 0
max (P)d
ea >
min (Q)
Since min (P)ea 2 ≤ V ≤ max (P)ea 2 , using Theorem 4.18 of [41] it can be concluded that ea converges
exponentially to a ball with radius

max (P) max (P)d
= (24)
min (P) min (Q)
which is a global uniform ultimate bound on the estimation error [41].
Remark. This bound can be minimized by using the relaxation in [43] and solving the following optimization problem:
find P, L i , ik , i = 1, 2, . . . , m, j = 1, 2, . . . , m, k = 1, 2, . . . , p and maximize 1 , 2 , 3 subject to:
⎛ ⎛ ⎞⎞
A i − L i C j Mi 0 · · · 0
⎜ ⎜ −1 C 0 I ··· 0 ⎟ ⎟
⎜ ⎜ i j ⎟⎟
⎜ ⎜ ⎟⎟
⎜ ⎜ .. .. .. . . .. ⎟⎟
i j = H ⎜ P ⎜ . . . . . ⎟⎟
⎜ ⎜ ⎟⎟
⎜ ⎜ p−1 ⎟⎟
⎝ ⎝ −i C j 0 0 · · · I ⎠⎠
p
−i C j 0 0 ··· 0
P = PT > 0
ii > 0
2
ii + i j + ji < −3 I
m−1
∀ i, j : ∃z : wi (z)w j (z) 0
−P > −2 I
P > 1 I (25)
for all i=1, 2, . . . , m.
4.2.2. Case 2
A similar, though notably more conservative bound can be found in the case when z is a function of x̂. For the
simplicity of the notation, the computation is presented for the case when the measurement matrices are common for
all the rules. The error system is that in (19) and let the condition
⎛ ⎛ ⎞⎞
A i − L i C Mi 0 · · · 0
⎜ ⎜ −1 C 0 I ··· 0 ⎟ ⎟
⎜ ⎜ i ⎟⎟
⎜ ⎜ ⎟⎟
⎜ ⎜ .. .. .. . . .. ⎟⎟
H ⎜P ⎜ . . . . . ⎟⎟ < −2Q
⎜ ⎜ ⎟⎟
⎜ ⎜ p−1 ⎟⎟
⎝ ⎝ −i C 0 0 · · · I ⎠⎠
p
−i C 0 0 ··· 0
P = PT > 0
Q = QT > 0
m

(wi (z) − wi (ẑ))(Ai x + Bi u + Mi d + ai ) ≤ e

i=1

Q − 2 I P
>0 (26)
P I
hold. Using a quadratic Lyapunov function for the extended error vector, V = eaT Pea , we obtain
⎛ ⎛ ⎞⎞
A i − L i C Mi 0 · · · 0
⎜ ⎜ −1 C 0 I ··· 0 ⎟ ⎟
⎜ ⎜ i ⎟⎟
m ⎜ ⎜ ⎟ ⎟
⎜ ⎜ .. .. .. . . .. ⎟⎟
V̇ = wi (z)eaT H ⎜ P ⎜ . . . . ⎟
. ⎟⎟ ea
⎜ ⎜ ⎟
i=1 ⎜ ⎜ ⎟ ⎟
⎝ ⎝ −ip−1 C 0 0 · · · I ⎠⎠
p
−i C 0 0 ··· 0

m
T
+ wi (z)2eaT P(0 0 0 · · · d( p) )T
i=1

m
+2eaT P (wi (z) − wi (ẑ))(I 0 · · · 0)T ( Ai x + Bi u + Mi d)
i=1
≤ −2min (Q)ea 2 + 2max (P)ea 2 + 2max (P)ea d

≤ −2(1 − )(min (Q) − max (P))ea 2 − 2((min (Q) − max (P))ea 2 − max (P)ea d )
where ∈ (0, 1) is arbitrarily chosen and Q = Q T is a positive definite matrix such that (26) is satisfied. Then, V̇ is
negative definite if
(min (Q) − max (P))ea 2 − max (P)ea d > 0

max (P)d
ea >
(min (Q) − max (P))
Since min (P)ea 2 ≤ V ≤ max (P)ea 2 , using Theorem 4.18 of [41] it can be concluded that ea converges
exponentially to a ball with radius

max (P) max (P)d
= (27)
min (P) (min (Q) − max (P))
This bound can also be minimized using the conditions (25), together with the condition min (Q) > max (P).
5. Estimation of unmodeled dynamics
Consider now the problem of estimating the states of a fuzzy system in the presence of unmodeled dynamics, i.e.,
the fuzzy system is of the form

m
ẋ = wi (z)(Ai x + Bi u + Mi ( A i x + B i u + i ))
i=1
y = Cx (28)
where Ai , Bi , i = 1, 2, . . . , m are known and the matrices A i , B i and the vectors i , i = 1, 2, . . . , m are unknown,
but A i , i = 1, 2, . . . , m are bounded by a known bound max , max A i ≤ max . This corresponds to the situation
when part of the true dynamics is unmodeled. The goal is to determine sufficient conditions and to design an observer
that estimates x and also the constant matrices A i , B i and the vector i , i = 1, 2, . . . , m. Therefore, our goal is to
estimate the unknown dynamics.
For the simplicity of the computations, we present only the case when the measurement matrix is common for all
rules of the model.
Assumption 1. It is assumed that Mi , i = 1, 2, . . . , m have full column rank, and rank(C Mi ) = rank(Mi ), i =
1, 2, . . . , m.
(1) Case 1: Consider first the case when the scheduling vector does not depend on states to be estimated. An observer
of the following form is considered:

m
x̂˙ = wi (z)( Ai x̂ + Bi u + L i (y − ŷ) + Mi ( Â i x̂ + B̂ i u + ˆ i ))
i=1
ŷ = C x̂
˙ = f ( Â , w(z), x̂, y)
Â i i i
˙B̂ = g ( B̂ , w(z), x̂, y, u)
i i i
˙ˆ ˆ
i = h i ( i , w(z), x̂, y) (29)
where L i , i = 1, 2, . . . , m are the gain matrices for each rule, and the update laws f i , gi , h i , i = 1, 2, . . . , m will be
determined so that the estimation errors x − x̂, A i − Â i , B i − B̂ i , and i − ˆ i converge asymptotically to zero.
The error system when using the observer (29) can be expressed as
m
ė = wi (z)[( Ai − L i C + A i )e + Mi ( Ā i x̂ + B̄ i u + ¯ i )]
i=1
e y = Ce (30)
with Ā i = A i − Â i , B̄ i = B i − B̂ i , ¯ i = i − ˆ i .
Consider first the following part of the error expressed in (30):

m
e˙ =
wi (z)(Ai − L i C + A i )
e (31)
i=1
Since a bound on A i , i = 1, 2, . . . , m is known, i.e., max A i ≤ max , stability conditions for perturbed fuzzy
systems can be used to render (31) stable and to design the gain matrices L i [42]: find P = P T > 0, Q = Q T > 0,
L i , i = 1, 2, . . . , m so that
min (Q)
max ≤
max (P)
H(P(Ai − L i C)) ≤ −2Q (32)
Consider now a Lyapunov function of the form

m
m
m
(¯i ¯i )
T
V = eT Pe + tr( Ā Ti Ā i ) + tr( B̄ Ti B̄ i ) +
i=1 i=1 i=1
for the error system (30), so that P satisfies (32). Then,

n
V̇ = wi (z)eT [(Ai − L i C + A i )T P + P( Ai − L i C + A i )]e
i=1

m
m
+2eT P wi (z)Mi Ā i x̂ + 2eT P wi (z)Mi B̄ i u
i=1 i=1

m
m
˙ Ā ) − 2
m
m
˙T
tr( B̂˙ i B̄ i ) − 2
T T
+2eT P wi (z)Mi ¯ i − 2 tr( Â i i (ˆi ¯i )
i=1 i=1 i=1 i=1

n
m
˙ T Ā ))
= wi (z)eT G i e + 2 (tr(x̂eT P Mi wi (z) Ā i ) − tr( Â i i
i=1 i=1

m m
˙T
(tr(ueT P Mi wi (z) B̄ i ) − tr( B̂˙ i B̄ i )) + 2
T
+2 (eT P Mi wi (z) ¯ i − ˆ i ¯ i )
i=1 i=1

n
m
˙ T ) Ā )
= wi (z)eT G i e + 2 tr((x̂eT P Mi wi (z) − Â i i
i=1 i=1

m m
˙T
tr((ue P Mi wi (z) − B̂˙ i ) B̄ i ) + 2
T
+2 T
(eT Pwi (z) − ˆ i ) ¯ i
i=1 i=1
with G i = H(P( Ai − L i C + A i )).
Since V > 0 and from (32) G i < 0, for i = 1, 2, . . . , m, V̇ < 0 is rendered negative definite if tr((x̂eT P Mi wi (z) −
˙ ) Ā ) = 0, tr((ueT P M w (z) − B̂˙ T ) B̄ ) and eT P M w (z) − ˙ˆ T = 0, for i = 1, 2, . . . , m. The equations lead to
Â
T
i i i i i i i i i
the update laws:
˙ = w (z)M T Pex̂T
Âi i i
B̂˙ i = wi (z)MiT PeuT

˙ˆ
i = wi (z)Mi Pe
T
(33)
Note that in general e is not directly available. However, given Assumption 1, there exist matrices i , i = 1, 2, . . . , m
so that i C = MiT P: i = MiT PC † , where C † denotes the Moore–Penrose pseudoinverse of C.
Therefore, the update laws can be expressed as
˙ = w (z)M T PC † e x̂T
Â i i i y
B̂˙ i = wi (z)MiT PC † e y uT
˙ˆ
i = wi (z)Mi PC e y
T †
(34)
If all the rules are sufficiently excited, both the error system and the estimation error of the unknown matrices are
asymptotically stable. It can easily be seen that, assuming nonzero and varying x, u, the only invariant set of the error
system (30) is e = 0, Ā i = 0, B̄ i = 0 and ¯ i = 0. If wi (z), i = 1, 2, . . . , m are sufficiently smooth and the fuzzy
model is defined on a compact set of variables, then based on Barbalat’s lemma and LaSalle’s invariance principle, the
dynamics (30), together with the update laws above are asymptotically stable.
The results can be summarized as follows:
Theorem 8. The error dynamics (30) is asymptotically stable, if the update laws (34) are used and, furthermore, there
exist P = P T > 0, Q = Q T > 0, L i , i = 1, 2, . . . , m so that
P>0
H(P(Ai − L i C)) < −2Q i = 1, 2, . . . , m
min (Q)
max ≤ (35)
max (P)
(2) Case 2: Consider now the case when z depends on x̂. The error system (30) becomes

m
ė = wi (ẑ)[( Ai − L i C + A i )e + Mi ( Ā i x̂ + B̄ i u + ¯ i )]
i=1

m
+ (wi (z) − wi (ẑ)) · (Ai x + Bi u + Mi ( A i x + B i u + i ))
i=1
e y = Ce (36)
m
Under the assumption that i=1 (wi (z) − wi (ẑ))(Ai x + Bi u + Mi ( A i x + B i u + i )) ≤ e, combining the
conditions in Theorems 5 and 8, we get:
Corollary 2. The error system (36), together with the update laws
˙ = w (ẑ)M T PC † e x̂T
Â i i i y
B̂˙ i = wi (ẑ)MiT PC † e y uT
˙ˆ
i = wi (ẑ)Mi PC e y
T †
(37)
is asymptotically stable, if there exist P = P T > 0, Q = Q T > 0, L i , i = 1, 2, . . . , m so that
P>0
H(P( Ai − L i C)) < −Q i = 1, 2, . . . , m

m
(wi (z) − wi (ẑ)) · ( Ai x + Bi u + Mi ( A i x + B i u + i )) ≤ e
i=1

Q − 2 I P
>0 (38)
P I
Remark. Note that if the measurement matrix is different for each rule of the fuzzy model, the update laws for the
matricesmof the unknown dynamics can still
mbe expressed† as (33). Update laws similar to (34) can be derived if there
exist ( i=1 wi (z)Ci )† , ∀z (Case 1), and ( i=1 wi (ẑ)Ci ) , ∀ẑ (Case 2), i = 1, 2, . . . , m. In this case, the update laws
are
†
˙Â = w (z)M T P w (z)C
m
i i i i e x̂T
i y
i=1
†
˙B̂ = w (z)M T P w (z)C
m
i i i i i e y uT
i=1
†
˙ˆ
m
i = wi (z)M T
i P wi (z)Ci ey (39)
i=1
if the scheduling vector does not depend

† on the states to be estimated and
˙ = w (ẑ)M T P
m
Â i i i wi (ẑ)Ci e y x̂T
i=1
†

m
B̂˙ i = wi (ẑ)MiT P wi (ẑ)Ci e y uT
i=1
†
˙ˆ
m
i = wi (ẑ)Mi P wi (ẑ)Ci
T
ey (40)
i=1
if the scheduling vector depends on x̂ and the observer gains are given by (35) and
(38), respectively. Note that,
for this case, to have a unique solution, Assumption 1 has to be modified to rank( mj=1 w j (z)C j Mi ) = rank(Mi ),
i = 1, 2, . . . , m, ∀z.
Remark. The results can also be applied if the unknown matrices have slow dynamics, such that A˙ i 0, etc.
6. Related work
Our objective is to estimate the states and unknown inputs of a nonlinear system. In this paper we consider nonlinear
systems represented by TS fuzzy models, and unknown inputs that are polynomial functions of time or can be approx-
imated by polynomials in time. Note that our goal is to estimate the states of the given physically relevant model, and
not a transformation of the states. To our knowledge, no other work in the literature pertains to this class of problems.
Many existing adaptive observers [11–16,23,47–49] require that the system is or can be transformed into the observer
canonical form. By assuming a canonical form, the physical meaning of the state variables is lost.
In fault detection and isolation, observer based methods have widely been used (see [50] and the references therein).
However, these methods usually concern linear systems [51] and they do not estimate the faults. A method for TS
systems in descriptor form has been proposed in [39] to estimate the states in the presence of unknown inputs. This
method is based either on decoupling the unknown inputs or on attenuating their effects on the states. If decoupling is
possible, the states are correctly estimated, and single faults can be isolated by using a bank of observers. However,
the faults (unknown inputs) cannot be reconstructed. For the case when the decoupling of all the unknown inputs is
not possible, the authors of [39] also proposed a method to attenuate their effect on the states. However, the estimation
error is proportional with the norm of the unknown inputs. In the current paper we consider unknown inputs that are
polynomial in time, whose norm generally grows unbounded and therefore, the estimation error grows unlimited if the
attenuating method of [39] is used.
Several methods [31,32,48] exist for nonlinear systems composed of a known linear part and a Lipschitz nonlinearity
affected by unknown inputs. An estimation method has been provided in [26] for a more general class of nonlinear
systems, where the known part of the nonlinearity is Lipschitz in the states and inputs. However, this method relies on
an assumption (Assumption 2 in [26]) related to a rank condition on a matrix composed of the direct feedthrough term
and of a distribution matrix of the unknown nonlinear terms affected by the unknown inputs. Since in this paper we
do not consider unknown inputs that affect the measurements directly, this assumption is generally not satisfied, and
therefore the method of [26] cannot estimate the unknown inputs.
For SISO LTI systems, the adaptive observer design has been largely investigated (see [10] and the references
therein). However, when applying adaptive state and input observers designed for linear systems, such as the approach
proposed in [27] for a nonlinear system, the observer can only be used in a small neighborhood of the linearization
point. Therefore, such observers are rarely able to estimate the states and the inputs of the nonlinear system. Moreover,
they can only deal with constant or slowly varying inputs.
7. Example
We illustrate the proposed design method on a simulation example. The system under consideration is the dynamic
model of a missile, adopted from [52], illustrated in Fig. 1. The nonlinear state-space equations are
˙ = K MCn (, , M) cos() + q
q̇ = K q M 2 Cm (, , M) (41)
Fig. 1. Schematic representation of a missile.
Table 1
Plant variables.
Angle of attack (deg)

q Pitch rate (deg/s)
M Mach number
c Commanded tail fin deflection (deg)
Actual tail fin deflection (deg)
Table 2
Plant parameters.
P0 = 973.3 Static pressure at 20 000 ft (lbs/ft2 )

S = 0.44 Surface area (ft2 )
m = 13.98 Mass (slugs)
vs = 1036.4 Speed of sound at 20 000 feet (ft/s)
d = 0.75 Diameter (ft)
Iy = 182.5 Pitch moment of inertia (slug ft2 )
a = 225
K = ( /180)0.7P0 S/(mvs )
K q = ( /180)0.7P0 Sd/I y
an = 0.000103 (deg−3 ) am = 0.000215 (deg−3 )
bn = −0.00945 (deg−2 ) bm = −0.0195 (deg−2 )
cn = −0.1696 (deg−1 ) cm = −0.051 (deg−1 )
dn = −0.034 (deg−1 ) dm = −0.206 (deg−1 )
The plant variables are given in Table 1, and the parameters in Table 2.
The Mach number M is an exogenous scheduling variable. The angle of attack and the pitch rate q are measured.
The actual tail fin deflection and the unknown inputs acting on the system have to be estimated. The aerodynamic
coefficients are expressed as
Cn (, , M) = sgn()[an ||3 + bn ||2 + cn (2 − M/3)||] + dn
Cm (, , M) = sgn()[am ||3 + bm ||2 + cm (−7 + 8M/3)||] + dm
The dynamics of the tail fin actuator are described by a first-order linear model:
˙= a + a c
0.8
μ (M)
0.6
0.4
0.2
0
2 2.5 3 3.5 4
M
Fig. 2. Membership functions for the Mach number M.
In order to use the proposed design, first an approximate TS fuzzy model of the system (41) is constructed. For the
Mach number M, five points M ∈ {2, 2.5, 3, 3.5, 4} and for the angle of attack seven points ∈ {−15, −10, −5, 0, 5,
10, 15} are chosen as centers of the membership functions. The membership functions for M are depicted in Fig. 2.
The scheduling vector consists of the Mach number M and the angle of attack , which is also a state to be estimated.
An example of a rule is:
If M is approximately 3 and is approximately 5 then ẋi = Ax + Bu + a, with
⎛ ⎞
−0.0012 1.0 0.0
⎜ ⎟
A = ⎝ −0.0445 0 −0.0399 ⎠
0 0 −225.0
B = (0 0 225)T
a = (0.0056 0.0839 0)T
where x = [ q ]T .
The affine consequent models in the fuzzy rules are obtained by Taylor series expansion around each combination
of the chosen M and . To compute the membership degree of the scheduling vector z = [M ]T , the algebraic product
operator is used.
To simulate the system, the input was randomly generated. For the considered estimations (unmodeled dynamics
and unknown input, respectively), the initial conditions for the nonlinear system were [15 2.3 3]T , while the initial
conditions used for the states and parameters to be estimated were zero. The LMIs for designing the observer in each
case were solved using the Yalmip toolbox [53]. The unknown input/unmodeled dynamics are assumed to affect and
q, therefore
⎛ ⎞
1 0
⎜ ⎟
Mi = ⎝ 0 1 ⎠
0 0
i = 1, 2, . . . , m.
(1) Adapting the TS model: Since the consequents in the fuzzy rules were obtained by linearizing the nonlinear
model, the local models represent the nonlinear system exactly only in the linearization points, i.e., in the combinations
of the points chosen for M and . When the system is not close to the linearization point or more than one membership
function is activated, the accuracy of the TS model decreases. This can be seen as if part of the actual dynamics
were unmodeled. In order to obtain a better approximation of the nonlinear system, first an observer to estimate this
“unmodeled dynamics” was designed, according to Theorem 8. This observer estimates the states and updates the
20
α
15 q
δ
State estimation error

10
−5
−10
0 0.2 0.4 0.6 0.8 1
Time [s]
Fig. 3. Estimation of unmodeled dynamics: state estimation error.
1 12
α d1
q 11 d
2
δ
0.5 10
Estimated input
0 8
−0.5 6
−1 4
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
Time [s] Time [s]
1 12
α d
1
q 11 d
2
δ
0.5 10
Estimated input
0 8
−0.5 6
−1 4
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
Time [s] Time [s]
Fig. 4. Estimation results for constant input. (a) Estimation error for the states using the proposed approach. (b) The unknown input (thin line) and
its estimate (thick line) using the proposed approach. (c) Estimation error for the states using the approach in [27]. (d) The unknown input (thin line)
and its estimate (thick line) using the approach in [27].
system matrices according to Eq. (34). The estimation error for the states is presented in Fig. 3. As can be seen, the
estimated states converge to the states of the nonlinear system. As a result of using this observer, for instance the rule
corresponding to the approximation in M = 3 and = 5 has been changed to:
1 12
α d
1
q 10 d
2
δ
0.5
Estimated input
6
0
4
2
−0.5
0
−1 −2
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
Time [s] Time [s]
1 12
α d
1
q 10 d2
δ
0.5
8
Estimated input
6
0
4
2
−0.5
0
−1 −2
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
Time [s] Time [s]
Fig. 5. Estimation results for second order input. (a) Estimation error for the states using the proposed approach. (b) The unknown input (thin line)
and its estimate (thick line) using the proposed approach. (c) Estimation error for the states using the approach in [27]. (d) The unknown input (thin
line) and its estimate (thick line) using the approach in [27].
If M is approximately 3 and is approximately 5 then ẋi = Ax + Bu + a, with

⎛ ⎞
0.0496 1.1488 0.0062
⎜ ⎟
A = ⎝ −0.0164 −0.0179 −0.0402 ⎠
0 0 −225.0
B = (−0.0032 − 0.0003 225.0)T
a = (0.00475 − 0.0031 0)T
(2) Unknown input estimation: Now assume that an unknown input d is acting on and q. Three cases are considered:
(1) constant input, (2) second-order polynomial input, and (3) a non-polynomial input. The application of the proposed
approach and the obtained results are detailed in what follows.
1. In the first case, d is constant, d = [5 10]T . The observer is designed based on the conditions of Theorem 6, with
the first derivative of the unknown input being already zero, i.e., correctly assuming a bias in the model. The state
estimation error is presented in Fig. 4(a). The unknown input and its estimate is presented in Fig. 4(b). Both the
estimated states and input converge to the correct ones, as expected.
1 4.5
α d1
q d
4 2
δ
0.5
Estimated input
3.5
0 3
2.5
−0.5
2
−1 1.5
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
Time [s] Time [s]
1 4.5
α d
1
q d2
4
δ
0.5
Estimated input
3.5
0 3
2.5
−0.5
2
−1 1.5
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
Time [s] Time [s]
Fig. 6. Estimation results for non-polynomial input. (a) Estimation error for the states using the proposed approach. (b) The unknown input (thin
line) and its estimate (thick line) using the proposed approach. (c) Estimation error for the states using the approach in [27]. (d) The unknown input
(thin line) and its estimate (thick line) using the approach in [27].
To compare with another method, we have implemented 1 the approach described in [27]. The state estimation error
and the estimated inputs are presented in Figs. 4(c) and (d), respectively. As can be seen, with the approach proposed
in this paper, the estimated states and inputs converge to their true values faster.
2. In the second case, d is a second order polynomial in time, d = [5 10]T t 2 + [3 1]T t + [2 4]T . The observer is
designed in the same manner as for the d in the constant input case, but using three derivatives in Theorem 6. The
estimation error for the states is presented in Fig. 5(a) and the unknown inputs and their estimates in Fig. 5(b). The
estimated states and inputs converge to the correct states and inputs as expected.
The state estimation error and the estimated inputs using the approach in [27] are presented in Figs. 5(c) and (d),
respectively. Similarly to the previous case, using the proposed approach, the estimated states and inputs converge
faster to their true values.
3. The last case considered is with the true unknown input d acting on the system being non-polynomial, given by
d = [1 3]T sin(t) + [4 2]T cos(t), but the observer is computed by (incorrectly) assuming a linear input, i.e., a ramp.
The estimation error for the states is presented in Fig. 6(a) and the unknown inputs and their estimate in Fig. 6(b).
The estimated input does not converge exponentially to the exact value of the true input (see Fig. 7), but only to a
1 We have also implemented the method described in [26]. However, due to the fact that Assumption 2 in [26] is not satisfied, only the states are
correctly estimated and not the unknown inputs. Therefore, those results are not presented here.
0.2
d1
0.15 d
2
0.1
Input estimation error

0.05
−0.05
−0.1
−0.15
−0.2
0 0.2 0.4 0.6 0.8
Time [s]
Fig. 7. Input estimation error using the proposed approach.
ball around it. The ultimate bound on the error (state and input), computed for this observer is 38.6 · d/ with
d the norm of the input and arbitrarily chosen in (0,1). However, based on the estimation errors presented in
Figs. 6(a) and 7, one can conclude that the computed bound is rather conservative.
The state estimation error and the estimated inputs using the approach in [27] for this case also are presented in
Figs. 6(c) and (d), respectively. As in the previous cases, using the approach proposed in this paper, the estimates
converge faster to their true values.
8. Conclusions
In this paper a method was proposed for estimating the state and unknown inputs of TS fuzzy systems. The design
of unknown input observers is important in practice, since in many cases not all the inputs are known. These unknown
inputs can represent disturbances acting on the process, effects of unmodeled dynamics, or actuator failures. The
proposed observer was designed based on the known part of the fuzzy model. Sufficient conditions were given for the
stability of the observer and the computation of the observer gains is based on solving a system of LMIs. Conditions on
the exponential convergence of the observer in the case of polynomial unknown inputs were presented and the design
guarantees an ultimate bound on the error signal. In the case of estimating unmodeled dynamics, sufficient conditions
were given for the asymptotic convergence of the observer. The design methods were illustrated on an example of
a missile. The simulation results showed that the proposed observer is able to estimate both the states and inputs
simultaneously.
In our future research, we will investigate the extension of the proposed method to stable state-feedback controller
design and also adaptive observer-based stable control design for practical control applications. Two main challenges
that have to be solved in this context are: the controller has to compensate for the unknown input that is time-varying;
moreover, if the scheduling vector depends on states that have to be estimated, the separation principle does not hold.
We will also investigate conditions under which fuzzy observers can be used to estimate other than polynomial types
of unknown inputs.
Acknowledgments
This research is sponsored by Senter, Ministry of Economic Affairs of the Netherlands within the project Inter-
active Collaborative Information Systems (Grant no. BSIK03024), and by the International Campus on Safety and
Intermodality in Transportation the Nord-Pas-de-Calais Region, the European Community, the Regional Delegation
for Research and Technology, the French Ministry of Higher Education and Research, and the French National Center
for Scientific Research.
References
[1] T. Takagi, M. Sugeno, Fuzzy identification of systems and its applications to modeling and control, IEEE Transactions on Systems, Man, and
Cybernetics 15 (1) (1985) 116–132.
[2] C. Fantuzzi, R. Rovatti, On the approximation capabilities of the homogeneous Takagi-Sugeno model, in: Proc. Fifth IEEE Internat. Conf. on
Fuzzy Systems, New Orleans, LA, 1996, pp. 1067–1072.
[3] K. Tanaka, T. Ikeda, H. Wang, Fuzzy regulators and fuzzy observers: relaxed stability conditions and LMI-based designs, IEEE Transactions
on Fuzzy Systems 6 (2) (1998) 250–265.
[4] M. Johansson, A. Rantzer, K. Arzen, Piecewise quadratic stability of fuzzy systems, IEEE Transactions on Fuzzy Systems 7 (6) (1999)
713–722.
[5] P. Bergsten, R. Palm, D. Driankov, Fuzzy observers, in: Proc. 10th IEEE Internat. Conf. on Fuzzy Systems, Vol. 2, Melbourne, Australia, 2001,
pp. 700–703.
[6] G. Feng, A survey on analysis and design of model-based fuzzy control systems, IEEE Transactions on Fuzzy Systems 14 (5) (2006) 676–697.
[7] K. Tanaka, H. Wang, Fuzzy regulators and fuzzy observers: a linear matrix inequality approach, in: Proc. 36th IEEE Conf. on Decision and
Control, Vol. 2, San Diego, California, 1997, pp. 1315–1320.
[8] P. Bergsten, R. Palm, D. Driankov, Observers for Takagi–Sugeno fuzzy systems, IEEE Transactions on Systems, Man and Cybernetics, Part B
32 (1) (2002) 114–121.
[9] R. Palm, P. Bergsten, Sliding mode observer for a Takagi-Sugeno fuzzy system, in: Proc. 9th IEEE Internat. Conf. on Fuzzy Systems, Vol. 2,
San Antonio, Texas, 2000, pp. 665–670.
[10] K.S. Narendra, A.M. Annaswamy, Stable Adaptive Systems, Prentice-Hall, Inc., Upper Saddle River, NJ, USA, 1989.
[11] J.-H. Park, G.-T. Park, Adaptive fuzzy observer with minimal dynamic order for uncertain nonlinear systems, IEE Proceedings—Control Theory
and Applications 150 (2) (2003) 189–197.
[12] S. Tong, H.-X. Li, W. Wang, Observer-based adaptive fuzzy control for SISO nonlinear systems, Fuzzy Sets and Systems 148 (3) (2004)
355–376.
[13] J.-H. Park, G.-T. Park, S.-H. Kim, C.-J. Moon, Output-feedback control of uncertain nonlinear systems using a self-structuring adaptive fuzzy
observer, Fuzzy Sets and Systems 151 (1) (2005) 21–42.
[14] H. Ho, Y. Wong, A. Rad, W. Lo, State observer based indirect adaptive fuzzy tracking control, Simulation Modelling Practice and Theory 7
(13) (2005) 646–663.
[15] Y. Wang, T. Chai, Output-feedback control of uncertain nonlinear systems using adaptive fuzzy observer, in: Proc. American Control Conf.,
Vol. 4, Portland, Oregon, 2005, pp. 2613–2618.
[16] J.-H. Park, P.-S. Yoon, G.-T. Park, Robust adaptive observer using fuzzy systems for uncertain nonlinear systems, in: Proc. 10th IEEE Internat.
Conf. on Fuzzy Systems, Vol. 2, Melbourne, Australia, 2001, pp. 749–752.
[17] S. Labiod, T.M. Guerra, Adaptive fuzzy control of a class of SISO nonaffine nonlinear systems, Fuzzy Sets and Systems 158 (10) (2007)
1126–1137.
[18] S. Tong, W. Wang, L. Qu, Decentralized robust control for uncertain T–S fuzzy large-scale systems with time-delay, International Journal of
Innovative Computing, Information and Control 3 (3) (2007) 657–672.
[19] S. Tong, H.-H. Li, Observer-based robust fuzzy control of nonlinear systems with parametric uncertainties, Fuzzy Sets and Systems 131 (2)
(2002) 165–184.
[20] M. Chadli, A. El Hajjaji, Comment on “Observer-based robust fuzzy control of nonlinear systems with parametric uncertainties”, Fuzzy Sets
and Systems 157 (9) (2006) 1276–1281.
[21] Q. Zhang, A. Xu, Implicit adaptive observers for a class of nonlinear systems, in: Proc. American Control Conf., Vol. 2, Arlington, Virginia,
2001, pp. 1551–1556.
[22] Y. Zhu, P. Pagilla, Adaptive controller and observer design for a class of nonlinear systems, in: Proc. 42nd IEEE Conf. on Decision and Control,
Vol. 2, Maui, Hawaii, 2003, pp. 1687–1692.
[23] W.-J. Wang, L. Luoh, Stability and stabilization of fuzzy large-scale systems, IEEE Transactions on Fuzzy Systems 12 (3) (2004) 309–315.
[24] J. Ruiz Vargas, E. Hemerly, Adaptive observers for unknown general nonlinear systems, IEEE Transactions on Systems, Man and Cybernetics,
Part B 31 (5) (2001) 683–690.
[25] N. Hovakimyan, A. Calise, V. Madyastha, An adaptive observer design methodology for bounded nonlinear processes, in: Proc. 41st IEEE
Conf. on Decision and Control, Vol. 4, Las Vegas, Nevada, 2002, pp. 4700–4705.
[26] Q.P. Ha, H. Trinh, State and input simultaneous estimation for a class of nonlinear systems, Automatica 40 (10) (2004) 1779–1785.
[27] Y. Xiong, M. Saif, Unknown disturbance inputs estimation based on a state functional observer design, Automatica 39 (8) (2003) 1389–1398.
[28] S. Zhou, J. Lam, A. Xue, H∞ filtering of discrete-time fuzzy systems via basis-dependent Lyapunov function approach, Fuzzy Sets and Systems
158 (2) (2007) 180–193.
[29] C.-S. Tseng, B.-S. Chen, H∞ decentralized fuzzy model reference tracking control design for nonlinear interconnected systems, IEEE
Transactions on Fuzzy Systems 9 (6) (2001) 795–809.
[30] C.-S. Tseng, B.-S. Chen, H∞ fuzzy estimation for a class of nonlinear discrete-time dynamic systems, IEEE Transactions on Signal Processing
49 (11) (2001) 2605–2619.
[31] A. Pertew, H. Marquez, Q. Zhao, H∞ synthesis of unknown input observers for non-linear Lipschitz systems, International Journal of Control
78 (15) (2005) 1155–1165.
[32] A. Pertew, H. Marquez, Q. Zhao, H∞ observer design for Lipschitz nonlinear systems, IEEE Transactions on Automatic Control 51 (7) (2006)
1211–1216.
[33] M. Corless, J. Tu, State and input estimation for a class of uncertain systems, Automatica 34 (6) (1998) 757–764.
[34] A.L. Sheldrake, Estimation of Plant Electrical Load (Chapter 1) in Handbook of Electrical Engineering: For Practitioners in the Oil, Gas and
Petrochemical Industry, John Wiley and Sons, New England, 2005, pp. 1–18.
[35] D.D. Li, C. Chen, A novel approach to estimate load factor of variable-speed wind turbines, IEEE Transactions on Power Systems 20 (2) (2005)
1186–1188.
[36] K. Guelton, S. Delprat, T.-M. Guerra, An alternative to inverse dynamics joint torques estimation in human stance based on a Takagi–Sugeno
unknown-inputs observer in the descriptor form, Control Engineering Practice 16 (12) (2008) 1414–1426.
[37] Y. Wang, M. Papageorgiou, Real-time freeway traffic state estimation based on extended Kalman filter: a general approach, Transportation
Research B 39 (2) (2005) 141–167.
[38] T.L. Liao, N.S. Huang, An observer based approach for chaotic synchronization with applications to secure communications, IEEE Transactions
on Circuits and Systems 46 (9) (1999) 1144–1150.
[39] B. Marx, D. Koenig, J. Ragot, Design of observers for Takagi–Sugeno descriptor systems with unknown inputs and application to fault diagnosis,
IET Control Theory & Applications 1 (5) (2007) 1487–1495.
[40] B. Kulcsar, J. Bokor, J. Shinar, Unknown input reconstruction for LPV systems, International Journal of Robust and Nonlinear Control 20 (5)
(2009) 579–595.
[41] H.K. Khalil, Nonlinear Systems, Prentice-Hall, Upper Saddle River, NJ, USA, 2002.
[42] P. Bergsten, Observers and controllers for Takagi–Sugeno fuzzy systems, Ph.D. Thesis, Örebro University, Sweden, 2001.
[43] H. Tuan, P. Apkarian, T. Narikiyo, Y. Yamamoto, Parameterized linear matrix inequality techniques in fuzzy control system design, IEEE
Transactions on Fuzzy Systems 9 (2) (2001) 324–332.
[44] M. Johansson, A. Rantzer, Computation of piecewise quadratic Lyapunov functions for hybrid systems, IEEE Transactions on Automatic
Control 43 (4) (1998) 555–559.
[45] M. Johansson, Piecewise linear control systems, Ph.D. Thesis, Department of Automatic Control, Lund Institute of Technology, Sweden, 1999.
[46] W.-J. Wang, C.-H. Sun, Relaxed stability and stabilization conditions for a TS fuzzy discrete system, Fuzzy Sets and Systems 156 (2) (2005)
208–225.
[47] R. Marino, P. Tomei, Adaptive observers with arbitrary exponential rate of convergence for nonlinear systems, IEEE Transactions on Automatic
Control 40 (7) (1995) 1300–1304.
[48] R. Marino, G. Santosuosso, P. Tomei, Robust adaptive observers for nonlinear systems with bounded disturbances, IEEE Transactions on
Automatic Control 46 (6) (2001) 967–972.
[49] C.-H. Hyun, J.-H. Kim, E. Kim, M. Park, Adaptive fuzzy observer based synchronization design and secure communications of chaotic systems,
Chaos, Solitons & Fractals 27 (4) (2006) 930–940.
[50] P.M. Frank, Fault diagnosis in dynamic systems using analytical and knowledge-based redundancy: a survey and some new results, Automatica
26 (3) (1990) 459–474.
[51] X. Zhang, T. Parisini, M.M. Polycarpou, Sensor bias fault isolation in a class of nonlinear systems, IEEE Transactions on Automatic Control
50 (3) (2005) 370–376.
[52] R. Nichols, R. Reichert, W. Rugh, Gain scheduling for H∞ controllers: a flight control example, IEEE Transactions on Control Systems
Technology 2 (2) (1993) 69–79.
[53] J. Löfberg,YALMIP: a toolbox for modeling and optimization in MATLAB, in: Proc. CACSD Conf., Taipei, Taiwan, 2004, pp. 284–289.

Lendek 2010

Uploaded by

Copyright:

Available Formats

Lendek 2010

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lendek 2010

Uploaded by

Copyright:

Available Formats

Fuzzy Sets and Systems 161 (2010) 2043 – 2065

Adaptive observers for TS fuzzy systems with

∗ Corresponding author. Tel.: +31 15 27 88573; fax: +31 15 27 86679.

2. Stability of fuzzy systems and observers

2.2. Autonomous fuzzy systems

Consider the autonomous fuzzy system:

Theorem 3. System (2) is exponentially stable if there exist matrices P = P T , Q = Q T , so that

2.3. Fuzzy observers

Consider now the affine fuzzy system

and an observer of the form

H(P( Ai − L i C j + A j − L j Ci )) < 0 ∀ j ≥ i : ∃ z : wi (z)w j (z) 0 (7)

2.3.2. The scheduling vector depends on the estimated states

observer (5) becomes

Consider the following fuzzy system:

4. Polynomial unknown inputs

4.1. Observer design

Consider the TS fuzzy system of the form

Theorem 6. The estimation error using the observer

which is negative definite if the condition (15) is satisfied.

The proof follows directly.

Then the extended error system becomes

The proof follows directly.

4.2. Bound on errors

if there exist P = P T > 0, Q = Q T > 0, L i , ik , i = 1, 2, . . . , m, k = 1, 2, . . . , p, j = 1, 2, . . . , m so that

≤ −2min (Q)ea 2 + 2max (P)ea d

which is a global uniform ultimate bound on the estimation error [41].

≤ −2min (Q)ea 2 + 2max (P)ea 2 + 2max (P)ea d

(min (Q) − max (P))ea 2 − max (P)ea d > 0

5. Estimation of unmodeled dynamics

Consider now a Lyapunov function of the form

for the error system (30), so that P satisfies (32). Then,

B̂˙ i = wi (z)MiT PeuT

The proof follows directly.

if the scheduling vector does not depend

Fig. 1. Schematic representation of a missile.

Angle of attack (deg)

P0 = 973.3 Static pressure at 20 000 ft (lbs/ft2 )

Fig. 2. Membership functions for the Mach number M.

State estimation error

Fig. 3. Estimation of unmodeled dynamics: state estimation error.

If M is approximately 3 and is approximately 5 then ẋi = Ax + Bu + a, with

Input estimation error

Fig. 7. Input estimation error using the proposed approach.

You might also like

if there exist P = P T > 0, Q = Q T > 0, L i , ik , i = 1, 2, . . . , m, k = 1, 2, . . . , p, j = 1, 2, . . . , m so that

(min (Q) − max (P))ea 2 − max (P)ea d > 0